DeepSeek V4 Flash: local coding tests on two DGX Sparks

This review tests the official DeepSeek V4 Flash July 31 edition as a local LLM in Open Code on two NVIDIA DGX Sparks. The presenter reports about 25 tokens per second during a coding task. A detailed cluster setup is reserved for a follow-up video.

The coding assistant builds an encyclopedia of the first 25 legendary Pokémon and an International Space Station tracker from prompts. The presenter describes the encyclopedia as functional but says its interface needs work. The tracker takes some back-and-forth, and its displayed location appears wrong at the end; the presenter suggests adjusting the API call.

The benchmark discussion covers DeepSeek's released results against earlier V4 models and GLM 5.2. The presenter cautions that coding scores use DeepSeek's own harness, so comparisons may not be directly equivalent. Internal DS-Bench results are also discussed, though those tests are not publicly available.

For hardware planning, the presenter gives model-loading estimates of about 168 GB of VRAM at 4-bit precision and 110 GB at 3-bit, with more memory needed for context. The stated context limit is one million tokens. Dwarf Star is discussed as another inference option using SSD offloading and mixed precision. DeepSeek V4 Flash is text-only in this review, without image or vision input.